The performance of Internet applications is heavily affected by the end-to-end available bandwidth. Thus, it is very important to examine how to accurately predict the available Internet bandwidth. A number of available bandwidth prediction algorithms have been proposed to date, but none of the existing solutions are able to achieve a high level of accuracy. In this paper, a Multi-manifold based Available Bandwidth prediction algorithm (MD-AVB) is proposed, based on the observation that the available bandwidth space on the Internet is multi-manifold and asymmetrical. In the proposed algorithm, the available bandwidth space is divided into multiple lower-dimensional domains iteratively, and each domain is embedded separately to predict the available bandwidth. Experiments on HP $S^{\mathrm{3}}$ datasets demonstrate that the proposed algorithm is more accurate than existing approaches.